Cognitive Monoculture

In the AI Story

The homogenization is structural rather than intentional. A large language model produces output that reflects the distribution of the text it was trained on. It generates code in the most common styles, using the most common frameworks, following the most common conventions. It does not produce the idiosyncratic approaches that individual practitioners develop through years of independent exploration — the unconventional solution that works for reasons no one anticipated, the heterodox methodology that fails in most contexts but succeeds brilliantly in the specific context for which it was developed. The model produces the mean. A sophisticated, high-quality mean — but a mean.

The history of science provides the evidence for why this matters. Breakthroughs that transformed fields came disproportionately from practitioners working outside the mainstream. Barbara McClintock's discovery of genetic transposition, dismissed for decades. Ignaz Semmelweis's insistence on handwashing, ridiculed before germ theory. Lynn Margulis's endosymbiotic theory, rejected before it reshaped evolutionary biology. Each breakthrough emerged from a cognitive niche the mainstream had not occupied. A cognitive monoculture reduces the probability of such breakthroughs by reducing the diversity of perspectives from which they emerge.

The monoculture also redistributes power. When all practitioners depend on the same tools provided by the same institutions, cognitive authority shifts from the distributed network of individual minds to the centralized corporations that train and maintain the models. The diversity of power — thousands of practitioners each constituting an autonomous node — is replaced by concentration in the institutions that shape the patterns of thought. Næss identified concentration of power as a structural threat to ecological health; the cognitive case extends this diagnosis into the domain of mind.

The preservation of cognitive biodiversity requires intervention that the market will not provide, because the market rewards the monoculture. The practitioner who maintains friction-rich practices when competitors have adopted AI-assisted production will produce less output, at least in the short term, and the market will not compensate her for the long-term resilience that her cognitive diversity contributes to the community. This is not market failure — it is a feature of a system that treats exchange value as the measure of all value. The deep-ecological response is to distinguish exchange value from ecological value and to build institutions that preserve what the market cannot price.

Origin

The concept extends Næss's third platform principle into cognitive ecology. It draws on agricultural monoculture's well-documented fragility patterns (Irish Potato Famine, Gros Michel banana collapse, Cavendish succession, industrial corn susceptibilities) and on the emerging literature on model homogenization and output convergence in large language models.

Key Ideas

Structural, not intentional. Homogenization is a feature of tool design, not a conspiracy of the users.

Output reflects the mean. LLMs produce the statistical central tendency of their training data; practitioners using them converge toward the same tendency.

Diversity is the immune response. Ecosystems survive disturbance because they can respond in multiple ways; monocultures have no redundancy.

Breakthroughs come from niches. The history of innovation is the history of perspectives that the statistical mean did not include.

Concentration of power. Cognitive monoculture centralizes authority in the institutions that shape the tools — a structural feature, not an accident.